ins

app.py

@@ -3,6 +3,7 @@ import gradio as gr
 from typo_check import css, process_input,MODEL_OPTIONS_TYPO
 from title_gen import generate_title, MODEL_OPTIONS_TITLE
 from content_gen import generate_content, MODEL_OPTIONS_CONTENT, get_default_prompt
+from instruct_dv import generate_response, MODEL_OPTIONS_INSTRUCT 
 
 
 # Create Gradio interface using the latest syntax
@@ -198,6 +199,42 @@ This is an experimental model trained on a very small dataset of Dhivehi news ar
 **Notice:**
 
 All outputs generated are synthetic, created using fine-tuned models for experimental and educational evaluation. Accuracy is not guaranteed, and the content should not be considered a source of truth. Please avoid applying these results to production environments, critical systems, or real-world decision-making without proper validation.
+""")
+
+        with gr.Tab("Instruction Following"):
+            gr.Markdown("# <center>Dhivehi Instruction Following</center>")
+            gr.Markdown("Enter an instruction and (optionally) input text. The model will generate a response following your instruction. Set the random seed for reproducibility. Enable sampling for creative/random results.")
+            with gr.Row():
+                instruction = gr.Textbox(lines=2, label="Instruction", rtl=True, elem_classes="textbox1")
+            with gr.Row():
+                input_text = gr.Textbox(lines=2, label="Input Text (optional)", rtl=True, elem_classes="textbox1")
+            with gr.Row():
+                model_choice = gr.Dropdown(choices=list(MODEL_OPTIONS_INSTRUCT.keys()), value=list(MODEL_OPTIONS_INSTRUCT.keys())[0], label="Model")
+            with gr.Row():
+                seed = gr.Slider(0, 10000, value=42, step=1, label="Random Seed")
+                use_sampling = gr.Checkbox(label="Use Sampling (Creative/Random)", value=True)
+            with gr.Row():
+                generated_response = gr.Textbox(label="Model Response", rtl=True, elem_classes="textbox1")
+            generate_btn = gr.Button("Generate Response")
+            generate_btn.click(
+                fn=generate_response,
+                inputs=[instruction, input_text, seed, use_sampling, model_choice],
+                outputs=generated_response
+            )
+            gr.Examples(
+                examples=[
+                    ["ދީފައިވާ މައުޟޫޢާ ބެހޭގޮތުން ކުރު ޕެރެގްރާފެއް ލިޔެލާށެވެ.","އިއާދަކުރަނިވި ހަކަތަ ބޭނުންކުރުމުގެ މުހިންމުކަން"],
+                    ["އާ މޯބައިލް އެޕް ޕްރޮމޯޓް ކުރުމަށް މާކެޓިންގ ސްޓްރެޓެޖީތަކުގެ ލިސްޓެއް އުފެއްދުން.",""],
+                    ["ދިގުމިނުގައި 10ސެންޓިމީޓަރު އަދި ފުޅާމިނަކީ 5ސެންޓިމީޓަރު ހުންނަ ރެކްޓަންގްލަރެއްގެ ސަރަހައްދު ހިސާބުކުރުން.",""],
+                    ["ތިރީގައިވާ ބަސްފުޅު ތެދެއް ނުވަތަ ދޮގުގެ ގޮތުގައި ގިންތިކުރުން.","ސުޕްރީމް ކޯޓަކީ އެމެރިކާގެ އެންމެ މަތީ ކޯޓެވެ."],
+                ],
+                inputs=[instruction, input_text],
+            )
+            gr.Markdown("""\
+**Notes:**
+- This tab allows you to give instructions to the model, optionally with input text, for general-purpose generation or task following in Dhivehi.
+- Try different seeds or enable sampling for more creative outputs.
+- The model is experimental and may not always follow instructions perfectly.
 """)
 
 # Launch the app

instruct_dv.py

@@ -0,0 +1,77 @@
+import random
+import numpy as np
+import torch
+from transformers import T5Tokenizer, T5ForConditionalGeneration
+import spaces
+
+
+# Available models
+MODEL_OPTIONS_INSTRUCT = {
+    "A2 Model": "alakxender/flan-t5-base-alpaca-dv5",
+    "A1 Model": "alakxender/flan-t5-base-alpaca-dv"
+}
+
+# Cache for loaded models/tokenizers
+MODEL_CACHE = {}
+
+def get_model_and_tokenizer(model_dir):
+    if model_dir not in MODEL_CACHE:
+        print(f"Loading model: {model_dir}")
+        tokenizer = T5Tokenizer.from_pretrained(model_dir)
+        model = T5ForConditionalGeneration.from_pretrained(model_dir)
+        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        print(f"Moving model to device: {device}")
+        model.to(device)
+        MODEL_CACHE[model_dir] = (tokenizer, model)
+    return MODEL_CACHE[model_dir]
+
+max_input_length = 256
+max_output_length = 256
+
+@spaces.GPU()
+def generate_response(instruction, input_text, seed, use_sampling, model_choice):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(seed)
+
+    model_dir = MODEL_OPTIONS_INSTRUCT[model_choice]
+    tokenizer, model = get_model_and_tokenizer(model_dir)
+
+    combined_input = f"{instruction.strip()} {input_text.strip()}" if input_text else instruction.strip()
+    inputs = tokenizer(
+        combined_input,
+        return_tensors="pt",
+        truncation=True,
+        max_length=max_input_length
+    )
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    inputs = {k: v.to(device) for k, v in inputs.items()}
+
+    gen_kwargs = {
+        "input_ids": inputs["input_ids"],
+        "attention_mask": inputs["attention_mask"],
+        "max_length": max_output_length,
+        "no_repeat_ngram_size": 3,
+        "repetition_penalty": 1.5,
+    }
+
+    if use_sampling:
+        gen_kwargs.update({
+            "do_sample": True,
+            "temperature": 0.1,
+            "num_return_sequences": 1,
+            "num_beams": 1,
+        })
+    else:
+        gen_kwargs.update({
+            "num_beams": 8,
+            "do_sample": False,
+            "early_stopping": True,
+        })
+
+    with torch.no_grad():
+        outputs = model.generate(**gen_kwargs)
+    decoded_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    return decoded_output